I study how people and models think together, and build it into real products.
I research how frontier models reason, remember and act, and how that fits with the way people actually think. Then I build it, and write down exactly how, so you can too. Most of it was first tested in regulated finance, where a wrong answer costs real money.
Frontier deployment patterns
Every pattern here came out of a real deployment, and none of them depend on one domain. Most of the evidence comes from regulated finance, the hardest place I know to ship an AI system: decisions carry consequences, facts get contested and every action needs an owner.
Enterprise Agent Fabric
A layered estate of foundation, domain and process agents that lets any large enterprise assemble governed agentic systems from certified parts instead of building each one from scratch.
View case study →Identity Resolution Under Ambiguity
An autonomous multi-agent investigation, proven on financial-crime screening, that works the ambiguous residual of deterministic matching and feeds a deterministic decision gate, with every disposition made by a human investigator.
View case study →Enterprise Knowledge and Context Layer
The shared context backbone every agentic use case in an enterprise stands on, from governed ingestion through intent-routed retrieval to a context envelope with full traceability.
View case study →Start reading here
Six pieces that carry the argument. Three on what frontier models are actually doing; three on turning them into systems a team can depend on.
Frontier research
The Reasoning Stack, 2026
Chain-of-thought, tool use, search, RL with verifiable rewards and test-time compute, ranked by evidence: what holds up and where the seams are.
Frontier researchWhat Scaling Laws Don't Tell You
Scaling laws predict loss, not capability. Four blind spots — data quality, post-training returns, architecture bets and evaluation ceilings — matter more to what ships.
Frontier researchAgent Research 2023–2026: What Held Up in Production
An audit of the agent literature since ReAct: which ideas reached production systems, which stayed benchmark artefacts, and why.
Forward-deployed engineering
The Eval Pipeline is the Product
Prompts are disposable and models are rented. The durable asset is the evaluation harness that tells you, before your users do, whether the system still works.
Forward deployedMulti-Agent Topologies with Computable Blast Radius
Treat agent tool permissions as a reachability graph and compute, before deployment, the maximum damage any single agent could cause.
Forward deployedCalibrated Cascade Routing in Production
Small models first, confident escalation: tiered cascades, calibrated confidence and semantic caching that cut inference cost while holding quality.
Long-form programmes and reference
Machine Intelligence Canon
200 papers in dependency order, from mathematical foundations to agents, production controls and human judgement.
Enter the canon →Human–Machine Convergence
Learning systems, philosophies of mind, consciousness science and governed human–machine architectures.
Enter the collection →Books and study guides
Book-length studies with a recommended first read for each subject: AI systems, mathematics, mind and practice.
Open the library →Research Radar
Recent work organised by topic and year, with the AI research reading list and a typed topic map alongside.
Open the radar →